The controlled experiment most worth knowing is Galesic and Bosnjak (2009, Public Opinion Quarterly), who randomised the announced length of a web survey (10, 20 or 30 minutes) and tracked what happened. The longer the stated length, the fewer people started and the fewer finished. Worse, quality decayed within the questionnaire: answers to questions positioned later were faster, shorter in open text, and more uniform in grids (the straight-lining pattern). The thirtieth question does not get the same mind the third one did.
Wanous, Reichers and Hudy (1997, Journal of Applied Psychology) supply the other half of the case: across 17 studies and 7,682 people, single-item measures of overall job satisfaction correlated 0.67 (corrected) with full multi-item scales, rising to 0.72 against the best-constructed scales. A carefully chosen single question captures most of what a battery captures, at a fraction of the burden.
Porter, Whitcomb and Weitzer (2004, New Directions for Institutional Research) ran the closest thing to a controlled experiment on survey fatigue. Seniors sent a long survey almost back-to-back with a second one answered the second at 57%, against 67% for seniors who got only the second, 10 percentage points lower. In the same semester, invitations to two earlier surveys lowered response to a third from 68% to 46%, a drop of 22 points. The authors conclude that fatigue "may have the biggest impact on surveys conducted back-to-back", and that surveys "conducted in a previous semester may not affect response rates, or the impact may be minimal". They also note that more relevant surveys may cause less fatigue. Our reading is that burden clustering matters more than frequency in itself.
The IWE's design rests on that distinction. Twenty-six asks of under a minute each, spread across a year, is a different object from four 20-minute questionnaires, even though the four questionnaires demand more total minutes.
Rolstad, Adler and Rydén (2011, Value in Health) meta-analysed 20 studies of questionnaire length and response burden and found lower response rates for longer questionnaires, but warned that the result "should be interpreted with caution because it is impossible to separate the impact of content from length", and concluded that instrument choice should rest on content, not length per se. So a short survey of irrelevant questions is still a bad survey, and brevity alone buys nothing. We agree. The twelve questions are single items: the Work Ability Score, the standard self-rated health question, and questions the IWE wrote as paraphrases of single-item measures from the published literature. The rotation is designed around what each question measures (see the three questions page).
Put the strands together and a design falls out.
Three items per occasion keeps each check-in under a minute, which helps sustain participation cycle after cycle. That matters because participation is the asset everything else depends on. The measurement literature calls this a planned missing-data design: two anchor items every cycle plus one rotating item from a pool, a structure formalised by Raghunathan and Grizzle (1995), who found little was lost by giving each person only part of a questionnaire, and developed in the three-form design of Graham et al. (2006). Because the engine decides the rotation, the question a person is not asked never depends on how they would have answered it (Rhemtulla and Hancock, 2016).
Fortnightly spacing sits on the right side of Porter's clustering finding, because each ask is small and spaced. It also sets the temporal resolution, with a new organisation-level reading every two weeks, each with its own confidence interval. An annual survey gives one reading a year.
The guarded metric is completion. Low completion makes even organisation-level estimates noisy, so the engine reports the completion rate beside every figure.
Weekly asking buys little, because we expect the things these questions measure to move over weeks to months, and weekly asks add the kind of clustered burden Porter links to lower response. Monthly asking halves the trend resolution. Longer surveys less often is the configuration the length evidence argues against most directly, and it concentrates burden in exactly the way that produced Porter's 22-point drop. Fortnightly times three is not the only defensible design, but it is the one the trade-off points at, and we have yet to find a published evidence case for a different cadence. If one exists, we will link it here.
Why is your cadence what it is, and what evidence supports it? How many cycles before you will show me a trend, and what stops you showing one earlier? What happens to my dashboard when completion drops: does the number change its label, or just quietly get less true? How long is each ask when you time it?
The interactive demo shows the fortnightly engine on a simulated team, including what it refuses to show. The same engine is free for any employer at alltoogether.com; the same floors of five and ten appear in the open OWHS standard at openworkplacehealth.org; and if you are building measurement into your own product, that is what this site is for.